In partnership with
Good morning.
I found a rule in my own AI setup that nobody could explain, not even the AI that had been following that rule for months. That discovery led to a full audit, a genuinely bad day, and an uncomfortable realization about how much I had stopped checking my own work. Here is what actually happened.
First time reading? Get your own free subscription here.
AI INSIGHT
How to Build and Maintain AI Instructions
I found a rule buried in my newsletter's AI instructions that changed a section header depending on whether the person in the story was a man, a woman, or a group.
I have no idea who wrote it. I have no idea why. It had been quietly running for months, and neither I nor the AI following it could explain where it came from.
That turned out to be one of a dozen forgotten rules hiding across four different instruction documents, all supposedly working from the same playbook. Finding them took an afternoon. Figuring out why they had been allowed to just sit there took a lot longer.
None of it started with a calm afternoon of tidying up. It started with a genuinely bad day trying to produce that week's content and watching the results get worse with every attempt, until it felt like the whole system had rolled backward two months.
I was swearing at Claude, out loud and in the chat, more than once. Not mild irritation, real anger, the kind that shows up when something you built and trusted just stops working and you have no idea why. By the end of that day I was angry enough to stop everything and ask a harder question: why does this keep happening, and what exactly is telling the AI what to do here? That question is what led to going back and reading everything in full for the first time in months.
The four files were built with Claude Skills, a way to give Claude standing instructions that carry across every new conversation instead of re-explaining yourself each time. But you do not need to know that name, or use Claude specifically, for any of this to apply to you. If you have ever set up custom instructions in ChatGPT, saved preferences in Gemini, or built any kind of standing prompt you reuse, you already have some version of this same setup.
It took a full audit to see the actual shape of the problem. Four Skills, each edited independently over months, each carrying its own version of the same basic rules: how to write, what to avoid, when to link out, what never to invent. None of them agreed with each other anymore.
A few examples from that audit:
One Skill had quietly reversed a decision made two versions earlier, because nobody checked why the earlier version existed before changing it back.
One Skill was missing a rule every other Skill had, which meant an entire category of mistake was possible in exactly one place and nowhere else.
One Skill still pointed toward a step that had already been replaced by something better, because the replacement never made it back into that file.
None of these individual edits looked like a problem when Claude made each of them. Each seemed easy to justify, the kind of change that goes by without a second thought. That is exactly what makes this kind of drift dangerous. A big mistake gets noticed and fixed fast. A string of smaller, forgettable ones keeps running for months, stacking up, invisible until the day they finally add up to something big enough to blow up a day of work and make you furious.
That day is what turned this into an actual audit (and me writing this article), not a quick patch. The whole thing, read start to finish, not skimmed.
Fixing it took more than telling Claude to clean things up. It took reading every Skill in full, checking each rule against what had actually happened in past conversations, and building something new: a handful of shared documents holding the rules that mattered everywhere, with each Skill pointing back to them instead of repeating them.
Here is the part I am least proud of. I had not read those files carefully from the beginning either. Something about the fact that an AI had written them made me trust the words more than I would have trusted my own, and less carefully than I would have checked a person's work. I let a machine's confidence stand in for my own judgment - it’s AI afterall! That was the actual mistake underneath all the others, and it is the one I am still correcting.
You might not need to comb through four separate documents the way I did. But you should probably spend one afternoon with whatever standing instructions you have already built, and the willingness to read the whole thing again, out loud if it helps, and ask of every line: do I still agree with this, or did I just never get around to deleting it?
Want to try this yourself?
"Act as an editor reviewing my saved AI instructions for consistency and accuracy. This might be a Claude Skill, ChatGPT custom instructions, Gemini saved info, or any other standing setup you reuse across conversations. Here is what I currently have: [paste those instructions]. Read through everything and flag: any rule that contradicts another rule, any instruction that seems outdated or unexplained, and any place where the same guidance repeats in more than one spot. For each flag, tell me exactly what you found and why it might be a problem."
READER POLL
Have you ever gone back and read your own saved AI instructions?
WHERE TO GO NEXT
More on this topic, from sources worth your time:
How to Personalize and Set Up Claude -- a direct next step for setting up your own standing Claude preferences the right way from the start. [Swap for option 2 or 3 in production notes if Arnie prefers a different fit.]
What are skills? (Claude Help Center) -- Anthropic's own explanation of the feature this piece is built around, for anyone who wants the technical version.
Wispr Flow -- what I used to dictate most of the back-and-forth that turned into this piece. If you talk faster than you type, worth trying.
FROM OUR PARTNER
News for Everyday Americans!
A massive shift is happening in the American Media. The corporate elite news media has lost the trust of the American people. Half the American people believe national news organizations intend to mislead, misinform, and push their bias.
There is a better way! Sign up today for a FREE newsletter called The Flyover. Without the hidden agenda, slant, or bias, their talented team of editors dig through hundreds of sources and pull out the most important news of the day.
Advertising Disclosure: We evaluate all recommendations of products and services independently. Clicking on links provided on this page may result in AI for Daily Living earning compensation, which supports independent publishers like us.
